Remote sensing and probabilistic sampling based method for determining the carbon dioxide volume of a forest
A remote sensing and probabilistic sampling based method for determining carbon dioxide volume of a forest can correlate aerial data, such as LiDAR, CIR, and/or Hyperspectral data with actual sampled and measured ground data to facilitate obtainment, e.g., prediction, of an accurate forest inventory, and corresponding carbon dioxide volume thereof.
The remote sensing and probabilistic sampling based method for determining the carbon dioxide volume of a forest relates generally to analyzing remote sensing data, such as digital images and LiDAR data, to extract, classify, and analyze aggregate and individual features, such as trees in order to produce an accurate forest inventory. More particularly, the remote sensing and probabilistic sampling based method for determining the carbon dioxide volume of a forest relates to a method for producing an accurate determination of the amount of carbon dioxide volume per acre of the forest.
The importance of calculating carbon sequestration in forestland is important for a variety of reasons. Deforestation is estimated to be responsible for 20-25% of the world's greenhouse gas emissions (IPCC 2001), including carbon dioxide (CO2). The ability to calculate baseline forested carbon stocks as well as monitor change over time with statistical precision and accuracy is critical to understanding advancements or declines in sequestration efforts in the United States and around the world.
The probabilistic sampling forest inventory method of collecting quantifiable inputs for the accurate calculation of carbon stocks can provide a significant advancement in carbon sequestration measurement and monitoring of forested carbon stocks. This advancement is important in potentially reducing the cost of data input collection as well as the ability to systematically process the inputs for calculations of potentially billions of acres with statistical precision and accuracy unmatched heretofore.
SUMMARYAn embodiment of a remote sensing and probabilistic sampling based method for determining the carbon dioxide volume of a forest as described herein can generally comprise processing remote sensing data which is indicative of tree attribute information; defining a sampling frame within the remote sensing data; determining a field plot corresponding to said sampling frame and collecting field plot data therefrom, said field plot data comprising actual tree attribute information; creating a regression formula using the correlated tree match database and the remote sensing data from the sampling frame; generating a correlated model using the regression formula; applying the correlated model to all remote sensing data to create an accurate forest inventory; and determining the probabilistic carbon dioxide volume per acre of the forest.
The remote sensing data can comprise LiDAR data, digital images, and/or property boundary information, and the tree attribute information can be tree height, diameter, and/or species. The field plot data can be actual, measured tree attribute information. The field plot data can be obtained via precise physical measurements of trees on the ground, and the field plot can be matched to the sampling frame using, for example, a highly accurate Geographical Information System (“GIS”) to ensure that the sampling frame matches up with the field plot where the field data is measured.
Generation of the correlated model can further comprise verifying the accuracy and/or the quality of the correlated model. Verifying the accuracy of the correlated model can comprise comparing the field plot against the model prediction. Verifying the quality of the correlated model can comprise using quality control means including checking and ensuring that there are no calculation errors.
Basically, the remote sensing and probabilistic sampling based forest inventory method described herein can generally comprise the use of probabilistic sampling based methods to accurately capture forest inventory. The remote sensing data can be aerial data, such as the aforementioned LiDAR data and digital images, e.g., color infrared spectral (“CCIR”) photography, and/or multispectral photography. Also, hyperspectral data can be used instead of multispectral or CIR data. Via a sampling frame and corresponding field plot, the remote sensing data can be correlated with actual sampled and measured field data to obtain (predict) an accurate forest inventory. The resulting tree inventory can represent an empirical description of tree attributes, such as height, diameter breast height (“DBH”) and species for every tree within the selected inventory area. This data can be utilized to create an accurate forest inventory and then determine the carbon dioxide volume per acre of the forest.
A remote sensing and probabilistic sampling based method for determining the carbon dioxide volume of a forest is described in detail hereinafter. In the following description, for purposes of explanation, numerous specific details of exemplary embodiments are set forth in order to provide a thorough understanding of the remote sensing and probabilistic sampling based forest inventory method. However, it may be evident to one skilled in the art that the presently described method may be practiced without these specific details. The method can be most suitably performed using a computer system, e.g., a processor, storage media, input device, video display, and the like.
The terms “carbon” and “carbon volume” as used herein are used interchangeably with “carbon dioxide” and “carbon dioxide volume” and are used in the context of determining the amount of carbon, or carbon dioxide, which is sequestered in a subject forest.
Probabilistic Design-Conceptual Level
The probabilistic sampling method described herein is based upon remote sensing data that is blended with field plot data and used to create a correlated model, and is represented at the conceptual level in
Field Data Correlation
The field plot data 54 can be obtained via precise physical measurements of trees on the ground, wherein the field plot 60 is matched to the sampling frame (or vice-versa) using, for example, a highly accurate geographical information system (GIS) to ensure that the sampling frame matches up with the field plot 60 where the field data 54 is being measured. One, or multiple, sampling frames (which can be randomly selected) and corresponding field plots can be utilized. A set of correlated field plots can create a set of geo-referenced points, each attributed with a tree's data. Taken over a range of random samples, these plots 60 can be classified as unbiased by a forest biometrician. “Unbiased” sampling methodology is critical to a forest inventory in the same way that GAAP (Generally Accepted Accounting Principles) is critical in the analysis of financial performance of corporate entities. Any bias introduced in the sampling methodology makes all measurement results suspect.
The field plot data 54 can further comprise actual field attributes 62 and field plot measurement data 66. The field attributes can include tree attributes such as tree species, tree diameter and tree height, which can be used to create a correlated tree match database 64. The field data correlation process can include plot center location correction 68 to ensure the field plot 60 accurately corresponds to the associated sampling frame.
Correlated Model Generation
Referring more particularly to
A stand is a group of trees that, because of their similar age, condition, past management history, and/or soil characteristics, are logically managed together. Stratification (creating strata) is the process of aggregating the forest into units of reduced variability. Plots are small areas selected in stands, where field measurements are made. Plot tree-polygon data is the data about tree polygons created for stands that happen to be inside these plots.
The correlated model generation 56 can comprise correlating 74 the remote sensing data 52 and the field plot data 54 (via the correlated tree match database 64) in order to derive formulas 76, e.g., regression formulas, and associated coefficients, for tree species, tree height, and tree diameter. The correlated model 56 can then be applied to all remote sensing data 52 to produce an accurate probabilistic sampling based forest inventory 58. The resulting inventory 58 can represent an empirical description of tree attributes, such as species, height, and diameter breast height (“DBH”) for every tree within the selected inventory area. This also enables an accurate determination of the carbon dioxide volume 59 of the forest.
Generation of the correlated model 56 can further comprise verifying the accuracy 70 and/or the quality 72 of the correlated model 56. Model accuracy 70 can be verified by comparing the DBH, as well as the height, volume, and stem number values (as measured on the field plot 60) against what the model predicts these numbers should be for the exact same areas. Additionally, specially designed “control plots” could be used in a model verification process. Next, statistical estimates based on these two different plot inventory sets can be calculated and analyzed. Model quality 72 can be verified using quality control means, which can comprise procedures to check and ensure that there are no calculation errors in the models being utilized. In general, model quality can be related to model accuracy.
Overview and ExamplesLiDAR and multi spectral imagery, for example CIR photography, could be used separately, but in preferred embodiments are used together to identify and characterize individual tree crowns. Operationally, it is feasible to collect data for entire forests, and to construct digital maps with stand boundaries and with polygons representing tree crowns. Techniques to create crown polygons, assign species and impute tree sizes are the subject of many ongoing research efforts in Scandinavia, the United States, and elsewhere.
A starting point for sampling can be a map with crown polygons attributed with species and possibly LiDAR height; which can be used as a sampling frame for a statistically valid forest inventory 58. The sample design might assign the stands to strata, randomly select stands for sampling, and might randomly choose two or more map coordinate locations within the selected stand polygons to serve as plot centers (for sampling frames) to obtain field plot data to correlate to the sampling frames.
Fixed-area field plots 60 can be installed at these selected locations (sampling frames). Departures from conventional inventory procedures are that the plot 60 is centered as close as is technologically feasible to the pre-selected coordinates, and the plot 60 is stem-mapped. A fixed-area image plot is identified in the sample frame and co-located with the ground/field plot. The field trees and crown polygons are matched. Models are then fit, or calibrated, to predict what the individual crown polygons actually represent in terms of tree counts, species, DBH's and tree heights. These models can be approximately unbiased for basal area and tree count by species at the strata level. Field trees that are not matched to crown polygons are modeled separately. The models are applied to the sample frame so as to estimate the entire forest inventory 58; the predicted trees are summed by stand to create stand tables. This data further enables an accurate determination of the carbon dioxide volume 59 of the forest. The modeling strategies, and the methods for handling plot co-location, tree and crown polygon matching, and stand boundary overlap all present interesting challenges, which are addressed by the present method, are described in more detail below.
More particularly, embodiments of the correlation of the remote sensing data 52 and the field plot data 54 can further comprise one or more of the following steps:
-
- measuring the field plot center using a survey grade GPS device.
- saving the tree species, height and DBH information for all trees in the plot.
- measuring the distance to the plot center and azimuth.
- adding relative tree locations to the GPS locations, and displaying these absolute locations overlaid on aerial digital and LiDAR imagery.
- field crews correcting the field plot center location to a location that results in the best match between tree locations on the digital and/or LiDAR imagery and the locations measured in the field.
- using a tree recognition algorithm to detect objects, i.e., tree polygons, on the digital and/or LiDAR imagery-optimally (but not necessarily) these objects correspond to individual trees.
- calculating tree polygons attributes, LiDAR height estimates, area, color (on CIR imagery), and/or estimated tree species.
- tree polygon objects located in the plot areas are extracted from the data and used for the procedures described below, matching and/or statistical analysis.
- using automatic field tree matching to create a table in which measured field tree records are merged with tree polygon objects based upon geographic proximity.
- manually fixing the tree matching described above based upon interpreter estimate that current field tree is either contributing to some pixels of the tree polygon that was created or it is not visible from air because of a larger tree that contributed all or some pixels of the tree polygon.
- using statistical analysis for the data set of field trees, tree polygon objects and/or the relations created in the two preceding steps.
- the prediction estimates, e.g., the probabilities that tree polygons correspond to 0, 1, 2, 3 . . . trees; the probabilities for tree species for these trees; and the probabilities for DBH and height for these predicted trees
- the prediction can also estimate the number of trees “not seen,” i.e., which have no correlation with the tree polygons.
- applying these predictions, obtained using the data for field plot areas, on polygons over the entire inventory area (an example of probabilistic sampling based predictions is provided hereinafter).
- for predicted tree DBH and height values, using appropriate models to predict the volumes, and then aggregating these values to create a stand level inventory.
Referring back to
Model accuracy 70 can be verified by comparing the DBH, as well as the height, volume, and stem number values (as measured on the field plots) against what the model predicts these numbers should be for the exact same areas. Additionally, specially designed “control plots” could be used in a model verification process. Next, statistical estimates based on these two different plot inventory sets can be calculated and analyzed.
Model quality 72 can be verified using quality control means, which can comprise procedures to check and ensure that there are no calculation errors in the models being utilized. In general, model quality can be related to model accuracy.
Example of Probabilistic Sampling Based Predictions
The data contained in the tree match database 64 is used to predict the stem count, DBH, basal area, total height and volume estimates for all stands, which can be used to calculate the carbon dioxide volume of the forest. The estimates are based on probability theory and estimating the probabilities of several conditional events. These events (referred to herein as tree record sequences or TRS) are saved into the database. Separate DBH, height, and volume models are created for separate TRS events.
The following is an example of a TRS table corresponding to a single polygon:
The input variables used for analysis are polygon area (A) and the polygon height calculated from the LiDAR data (H). Also, species group prediction (S) was used, which was calculated from CIR imagery. For purposes of this example, only 2 species groups are used, namely—hardwood and conifer.
For all tree record sequence events, the following variables are calculated:
-
- E(C) or estimated count,
- the species group (SG),
- DBH,
- total height of the tree, and
- the volume of the tree.
These are output variables of the prediction equations. The following example illustrates a manner of predicting all tree sequence values for tree polygon objects, using a regression analysis:
-
- Estimated counts for TRS events are predicted for continuous values of A, H and the discrete values of S:
- Prediction formula for the probability Pr(C>=1) is estimated for continuous values of A, H and discrete values of S
- Prediction formula for the probability Pr(C>=2|C>=1), depending on A, H and S. The ‘|’ denotes the conditional probability.
- Prediction formula for the probability Pr (C>=3|C>=2), depending on A, H and S
- Estimated tree specie groups for TRS events depend on position of the tree (P). The largest tree of the polygon has the P value 1, the second largest 2 and so on. The probability of the largest tree specie group was
- Prediction formula for Pr{SG=1 for P=1|C=1]
- Prediction formula for Pr{SG=1 for P=2|C=2)
- Prediction formula for Pr{SG=1 for P=2|C=2, SG=1 for P=1)
- Prediction formula for Pr{SG=1 for P=2|C=2, SG=2 for P=1)
- Estimated counts for tertiary trees as follows:
- Prediction for E{C for P>=3|C=3}
- Estimated DBH values
- Prediction for DBH(A, H|SG=1, P=1)
- Prediction for DBH(A, H|SG=2, P=1)
- Prediction for DBH(A, H|SG=1, P=2)
- Prediction for DBH(A, H|SG=2, P=2)
- Estimated height values
- Prediction for HT(A, H|SG=1]
- Prediction for HT(A, H|SG=2]
- Volume equations
- Volume equations are not predicted. Instead, standard equations for the forest type are used to calculate volume from tree breast height diameter and total height values.
- Carbon weight equations
- Carbon Weight equations are also not predicted. Instead, Standard equations for species are used to calculate tree biomass, above ground carbon values, and root carbon values using previously determined tree breast height diameter and total height values.
Prediction Formula Example:
Pr{C=1}=1/(1+exp(c0+c1×A+c2×H+c3×A*H))
- Carbon Weight equations are also not predicted. Instead, Standard equations for species are used to calculate tree biomass, above ground carbon values, and root carbon values using previously determined tree breast height diameter and total height values.
- Estimated counts for TRS events are predicted for continuous values of A, H and the discrete values of S:
In this equation, the coefficients, c0 through c4, can be approximately 2.43, −0.0423, −0.0508 and, 0.00044, respectively.
Using the described predictions, estimated counts, diameters, heights, and volumes for all tree record sequences can be calculated. By summing up these results over all of the polygons in the stands, a more accurate stem count, basal area, and volume estimate for whole stands can be calculated.
Referring now to the diagrams in
Generally, the imagery data processing 103 can comprise polygon fusing and color 127, LiDAR processing 130, and CIR processing 133. The tree polygon classification 106 can generally comprise creating a training set 136, creating a sample plan 139, and creating tree crown polygon 142. The field data correlation 109 can generally comprise creating polygon match files 145, fixing the plot center 148, sample plot attributes 151. The correlated model generation 112 can basically comprise creating a species probability prediction model 154, a diameter probability prediction model 157 and the height probability prediction model 160. The probabilistic inventory creation 115/118 can generally comprise a plotting and regression formula to all tree crown data 163, providing an accuracy statement 166 and updating customers' original stand inventory data 169. This data can then be utilized to calculate the carbon dioxide volume 172 of the forest.
The determination of the carbon dioxide volume of the forest can generally involve applying known equations to the previously determined forest inventory data to determine the carbon dioxide volume per acre of the forest. The manner of determining the carbon dioxide volume can generally comprise applying known equations to convert the stem volume (in the case of above ground carbon volume) to carbon dioxide volume. Examples of physical data calculations to determine tree volume and corresponding carbon dioxide volume are:
Diameter breast height=(6.043*sgrt(A))+1.283.
Tree Height=div((A*A),((2.09000+(0.14050*A))*(2.09000+(0.14050*A))))+1.30000.
Stem Volume=00022927*pow(A,1.91505)*pow(0.99146,A)*pow(B,2.82541)*pow((B−1.3),−1.53547).
Above Ground Carbon Volume=exp(4.938+2.406*log(D))/2000000.
Root Carbon Volume=Root carbon=0.2*Above Ground Carbon Volume.
-
- (A=tree crown area, B=predicted tree height, and D=diameter breast height)
Accordingly, the Carbon Volume Equations Permit a Very Accurate Calculation of carbon dioxide volume per acre of the forest, based upon the forest inventory data acquired by the probabilistic sampling method.
Turning now to
As illustrated, the color infrared/stand line processing 172 can comprise multiple steps, including one or more of the following:
-
- A reading the input data 178;
- B splitting the CIR imagery and stand lines into smaller blocks, which can be saved in, for example, a split block data set 193 and smaller block files 196;
- C morphological opening and smoothing to create a smoothed block data set 199;
- D stand fixing/photo interpretation;
- E shape clipping, which can be saved as clipped shape files 202;
- F merging small blocks into one property file;
- G quality control and inheritance; and then
- H stratification, after which data can be saved as final property files 205.
Creating the smoothed block data set 199 can comprise rasterizing the stand boundaries to remove all possible topology errors and features below a certain size that may be present in original stand boundaries. Afterwards, the morphological opening can be applied to the rasterized stand map, followed by vectorizing the stand shape again, generalizing and smoothing the shape and finally clipping the boundaries to the property boundaries.
As further illustrated, the LiDAR processing 175 can also comprise multiple steps, including one or more of the following:
-
- I calculate DTM (digital terrain model) and saving import attributes 208;
- J select highest pixel and subtract DEM (digital elevation model);
- K map digital surface value to 8 bit integer;
- L convert and save data to gray scale, and the LiDAR metadata and attributes 211 can be saved, as can gray scale bitmap image files 214.
Generally, the LiDAR processing can comprise calculating the DEM; selecting highest pixel and subtracting DEM; mapping digital surface value; and converting the data to gray-scale. The laser scanning data provides a geo-referenced point cloud about earth's surface, DEM, which can include features like vegetation and buildings. The DTM can be calculated as a filtered version of DEM, and may contain only points classified as the “ground.” Both DEM and DTM values are then calculated for some grid (for example 0.5×0.5 meter grid or 1.5×1.5 foot grid). If more than one DEM point hits one grid cell, the highest one is selected. If no DEM points hit the grid, the values are interpolated using nearest points. A DSM (digital surface model) is then calculated as the difference between the DEM grid and DTM grid. After that, the continues values of DTM grid are replaced with discreet digital numbers between 0 and 255, and the results are saved as 8-bit grayscale bitmap files.
-
- A superimpose data sets and shift polygons, and a polygon shift dataset 229 can be saved;
- B manual review species strata, wherein a species strata excel report 232 and species strata data set 235 can be created;
- C calculate average color-infrared band for individual tree crowns;
- D calculate second order gray level texture feature;
- E selecting a subset of stands for classification, which can be saved as selected stands file 238;
- F create training set for species at strata level, which can be saved as a training data set per strata 241;
- G create classifier formula using discriminant analysis; and
- H batch classify polygons for all stands and strata, after which a classified polygon relational data store 244 can be created, as can be a shape file 246.
Generally, classifying tree polygons 106, in an embodiment thereof, can comprise creating a polygon shift dataset; calculating an average CMR, or multispectral, band for individual tree crowns; calculating a second order gray level texture feature; selecting a subset of stands for classification; creating a training set for species at strata level; creating a classifier formula using discriminant analysis; and batch classifying polygons for all stands and strata. Polygon shift refers to tree polygons created using the LiDAR data which are moved to locations where the same features are visible on the CIR or (multi-resolution) imagery. After shifting, average band values for these polygons are created for all CIR or (multi-resolution) image bands. Also, second order (spatial co-occurrence homogeneity) texture values are calculated for individual tree polygons. Based on stratification, some stands are selected for classification training. For these training stands, interpreters can select tree polygons and mark which species group they belong to, such as by using a computer mouse. The training set data (average band and texture values for the tree polygons classified by interpreter) are then analyzed using statistical software, and classes can be separated by a statistical method, such as a discriminant analysis. The resulting classifier is then applied for all stands and strata, and all tree polygons are assigned the probabilities of belonging to an identified species group.
-
- A measure plot center;
- B capture tree attributes, for example, height, location, segment, and others, which can be saved in a field tree attributes data set 266;
- C correct plot center;
- D field data quality control; and
- E create match data, wherein a field match relational data store 272 and polygon match files 275 can be saved.
-
- A sample stand data aggregation, and storing correlated aerial remote sensing and field info 293;
- B correlate strata, stand, plot, plot tree, plot tree polygon data, to create formulas and correlation coefficients, and storing such formulas and correlation coefficients 296;
- C species probability prediction process;
- D diameter probability prediction process;
- E height probability prediction process; and
- F a validation process for each of the three prediction process (for example, verification of accuracy and quality), which can include storing species probability prediction model forms and parameters, diameter probability prediction model forms and parameters, and height probability prediction model forms and parameters.
Moreover, as described previously, the correlated model generation 112, in an embodiment thereof, can comprise combining the data inputs listed above, which can include, for example, data output from one or more of the preceding processes, such as imagery data processing 103, tree polygon classification 106 and field data correlation 109. As also similarly described above in connection with
Basically, the correlated model generation 112 can comprise correlating strata, stand, plot, plot tree, and plot tree polygon data; and creating formulas to determine tree species; height; and diameter. Further processing can comprise verifying model accuracy and model quality to ensure an accurate ground/forest inventory is produced. An example of a probabilistic sampling based prediction is provided above in connection with
-
- A extract coefficients and regression formula, and storing coefficient and formula data set 320;
- B apply regression formula to all tree crown data, and storing probabilistic tree attributes at stand level 323;
- C calculate volume per acre, and store probabilistic volume per acre data set 326;
- D calculate stand summaries; and
- E update customer original stand data with inventory, which can include a final inventory, and storing accuracy statement regression formulas 329, a customer property file 332, and/or a relational stand inventory 335.
The probabilistic inventory generation 115, in an embodiment thereof, can generally overlap with the correlated model generation 56 described in connection with
In contrast, although similarly named, the correlated model generation 112 does not apply the formulas and coefficients created in that step and apply them to all the tree crown data. Instead, the process of extracting the formulas and coefficients and then applying them to all the tree crown data to create the forest inventory is performed in the probabilistic inventory generation step 115.
The process 118 can be performed instead of the process 115 in
-
- A extract coefficients and regression formula, and storing coefficient and formula data set 350;
- B apply regression formula to all tree crown data, and storing probabilistic tree attributes at stand level 353;
- C calculate volume per acre, and store probabilistic volume per acre data set 356;
- D calculate carbon volume per acre, and store probabilistic carbon volume per acre data set 359;
- E calculate stand summaries; and
- F update customer original stand data with inventory, which can include a final inventory, and storing accuracy statement regression formulas 362, a customer property file 365, and/or a relational stand inventory 368.
Accordingly, the probabilistic inventory generation process 118 for carbon dioxide volume determination can be performed to provide an accurate estimate of the carbon dioxide volume per acre of the forest.
A Method of Feature Identification and Analysis
A method of feature identification will now be described in connection with
The method can be most suitably performed using a computer system, e.g., a processor, storage media, input device, and video display in operable connection. Referring now to
The aggregate features of the digital image are separated into relatively homogeneous parts using a segmentation algorithm. In particular, a digital image of a portion of a forest, which typically depicts one or more species of trees of varying sizes, is segmented into stands of trees, which are preferably more homogeneous in composition than the forest itself. The stands are themselves segmented into polygons which encompass individual trees depicted in the portion of the digital image encompassed by the stand segment, such that the individual crowns of the trees are delineated by the polygons. The trees are then analyzed based on their crown area, classified by species or form model, or both, and using publicly obtainable forestry data and ratios to determine the tree's DBH and stem volume for each tree of a given species. The results of the classification and analysis are then compiled and saved in a searchable database, e.g., a vector file, such that a user of the system can determine the total stem volume for species of tree, or for trees of a range of DBH, or both, i.e., the total stem volume for a species of tree, including only trees of a certain minimal DBH, and optionally containing an accurate identification of the location and ownership of the trees, which is publicly available in tax parcel maps though difficult to obtain ordinarily. This information is particularly useful in the field of forestry, as it directly relates to the age of the forest the health of the forest, and economic value of the trees contained in the forest, particularly since the location of the economically valuable trees is also identified.
Typical digital images for use in this method are taken from aerial platforms or satellites and are either stored digitally when taken or transferred into digital format. As such, the input images contain digital numbers associated with pixels on the image. Typical sources for digital images include digital or film cameras or spectrometers carried by aircraft or satellite. At least visible color channels and infrared bandwidths can be used. Optionally, high pulse rate laser scanner data is used in combination with digital imagery. Digital input imagery is preferably of a resolution of 1 meter, more preferably 0.5 meter. Preferably, input images are ortho-rectified to a geo-coded map and color balanced.
High Level Segmentation
According to one aspect of the current invention, segmentation by a seeded region growing method is performed to obtain a segmentation vector file of polygon boundaries for homogenous areas within the digital image, e.g., tree stands. Referring now to
This describes the size of the discrete Gaussian filtering required to eliminate specific details and is preferably between 0 and 30.
If desired, a gradient image analysis is performed to identify homogenous areas within the input image. According to one embodiment of the method, gradient image analysis is performed by replacing the digital image with a new image corresponding to the greyscale gradient values of the image. A “seed point” is planted at the center of each region that has similar color/grayscale values. The similarity is measured in the gradient image, where a “postfilter” parameter specifies a gradient window size, where a window is the distance between the center and outer pixel that are selected by the algorithm to calculate the gradient. Preferable windows for segmentation of forested regions range from 1 to 30, preferably 15 depending on the resolution of the digital image and the separation of the trees imaged. The pixel with the lowest gradient is assigned the segment's seed point, and a homogenous region is grown from the seed points by adding pixels into the segments in the minimum change direction among all segments in the image. The added pixels must be next to an existing segment in any current phase. Adding pixels is continued until the entire image has been saturated according to the seeded region growing method and all pixels belong to a segment, and pixels at the borders of the segments represent the segment polygons. Boundary lines are drawn around the homogenous areas grown. Aggregate feature segmentation according to one embodiment is preferably performed on input images of high resolution, 0.4 to 1.5 m/pixel.
Accordingly, segment boundary lines, or polygons, are formed around the homogenous segments which are preferably polygonal in shape, as indicated in
Where first performed by an automatic or unsupervised algorithm, segmentation may preferably be adjusted using unsupervised and/or manual adjustment of the segmented image file. Referring once again to
Manual segmentation adjustment is performed by user selection of two or more neighboring segment polygons by drawing a line touching segment polygon borders using a mouse or other computer pointing device. Alternatively, the user drawn line may be connected at the start and end points and all segments that have points common with the line or that lie within the center of the line with connected start and end points will be merged. Manual segmentation adjustment is indicated in
The resulting segmented image file is stored in a vector file and can be displayed as an overlay or layer on the input image using ordinary display means. The segmented boundaries are stored in vector file format, such that the resulting layer can be drawn onto the original input image and/or rectified into any map coordinate system.
According to another embodiment of the present invention, no segmentation is required and a known boundary around an area on an input image is used to further analyze features within the image.
Low Level Segmentation
According to one aspect of the current invention, low level segmentation, or individual feature segmentation is performed on a segment selected from the high level segmentation file. Referring to
Additionally, a seed threshold may be selected as the threshold value of a given local maximum corresponding to a RGB, luminance, or another color space, which is used as a seed point from which to begin growing the low level segment according to a seeded region growing algorithm. The seed threshold in 8 bit images is between 0 and 256, preferably between 30 and 100. Alternatively, the seed threshold is another color parameter. Optionally, a cut ratio may also be used to filter out features on the image that will be considered background and left outside the remaining segments or individual tree crowns. The cut ratio is a threshold greyscale value of background, using the lowest grayscale value in the used color space that should be included in the segments. Values lower than this cut ratio will be considered as background and left outside the growing segments. The cut ratio in 8 bit images is between 0 and 256, preferably between 30 and 100. Alternatively, the cut ratio is another color parameter.
According to one embodiment of the present invention, seed points are placed at local maxims on the image that are brighter than a given threshold value. The brightness can be measured in luminance value of the image or some of the channels, or in a greyscale layer created using channel transformation based on channel information such as the calculated distance from a given color value in RGB, hue, saturation, or luminance-space and using that as the new greyscale value. This method makes it possible to find trees of a certain color and exclude other trees that have different color. Beginning from the seed points, individual low level segments are grown by adding pixels into the segments in the minimum change direction in the given greyscale layer, among all segments in the image. The added pixels must be next to the existing segment in any current phase. Adding pixels is continued until the given cut ratio parameter value in the given greyscale space is achieved or the entire image has been saturated and all pixels belong to a segment. Boundary lines are drawn around each resulting segment, such as a delineated tree crown. Tree crown segments from low level segmentation are illustrated on
Low level segmentation by seeded region growing and controlled filtering is performed according to methods described in the above-referenced related copending patent application.
The resulting vector file containing low level segments, such as tree crowns, is displayed as an overlay using ordinary display means.
Classification
According to one aspect of the present invention, individual segments are classified according to species or class using supervised classification. Preferably, feature classification is performed on individual tree stands from a forestral digital image using a training procedure. Referring now to
Remaining unselected tree crowns that correspond to those feature values are recognized and classified accordingly, as shown in
As indicated in
The resulting classification information is stored in a vector file. Preferably, the species information is appended to the tree crown vector file.
Data Analysis
According to another aspect of the present invention, mathematical models are used to analyze additional attributes associated with segmented and classified features in aggregate segments. Preferably, classified tree crowns within a homogenous tree stand are used for analysis. Crown correlation models are based on data obtained through field measurements based on species specific variables including, but not limited to, actual field measurement of tree crown size, DBH, volume, form class, and height. Models are stored in a database or model file, e.g. in XML format. Table 1 illustrates the Norway Spruce Model, which may be commonly used in analysis of Appalachian regions:
Any model that is commonly known in the art may be used to generate species specific data based on tree crown area. Examples of models commonly known in the art include Naeslund and Laasasenaho2.
Any variable may be included in the model, including but not limited to species, DBH, form class, tree quality or value. Models may be updated and altered depending on the species and region, and new models incorporated at any time into the model database.
Referring now to
Diameter breast height=(6.043*sgrt(A))+1.283;
Tree Height=div((A*A),((2.09000+(0.14050*A))*(2.09000+(0.14050*A))))+1.30000; and
Stem Volume=000022927*pow(A,1.91505)*pow(0.99146,A)*pow(B,2.82541)*pow((B−0.3),−1.53547).
In the equation, A=tree crown area.
According to another embodiment of the present method, batch modeling of delineated and classified features is performed using pre-selected models.
As indicated in
According to one embodiment of the present method, statistical information is also generated based on the modeling results, which includes, but is not limited to valuation of timber, estimation of property values based on public tax and terrain slope information, over or under-valuation of property by comparing market value to timber valuation, and estimation of vegetation growth rates and agricultural production. For example, the value of timber in a particular tree stand is calculated using the tree stem volume calculated from crown area, and public market value information based on species. Total volume of species used for this determination can be limited depending on the size of the tree as specified by the user. The market value may be obtained from public information or may be user input.
Another example of valuation information that can be generated from a digital image is orchard output. For example, where crown areas are captured from a digital image of a grove of orange trees, an estimate of the oranges produced by the individual trees can be calculated, e.g., by applying an empirically based statistical classification model where crown areas of area A1 produce 01 oranges, A2 produce 02 oranges, where A(x) is a range of areas, and O(x) is average orange production for areas A(x).
Statistical data is stored in the corresponding crown and/or stand vector file as indicated in
It is recognized that the scope of the present method includes application of the current method to other empirical models that are based on species data, such as fruit and juice production from fruit baring trees, carbon production, etc and that the present method is not limited to any specific embodiment presented herein.
EXAMPLE 1A 2 foot digital ortho-rectified, color-balanced image in TIFF format was taken of a 12 square mile forested area in Nicholas County, W. Va. The image was taken in RGB true color, and was taken in the fall when leaves are in senescence. Stand segmentation was performed using seeded region growing. Tree crowns were captured using segmentation, and filtering parameters used to eliminate undesirable details were a prefilter value of 4, a seed threshold of 90, and a cut ratio of 90. Species were classified according to supervised classification based on the teaching method. Three trees were selected per species. Three species were selected and identified using nearest neighborhood classification: poplar, red maple and red oak.
For data analysis, a model was selected for each of the three species based on data from 200 field measurements of different sized trees in Pennsylvania. The resulting data was displayed and is illustrated in
A stand area of 24 acres was selected south of Dugway Rd, in Madison County, N.Y., Tax Map Number 148-1-7. Low level segmentation was performed to delineate tree crowns, and species classification and tree crown data analysis were performed to determine tree species and total tree stem volume in board-feet. A total of 93,402 board-feet was calculated based on only trees of DBH greater than 12 inches. Trees with DBH greater than 25 inches were not used in the data analysis.
Species classification resulted in 85% Hard Maple, 13% undefined, and 2% Chemy. The Norway Spruce Model was selected based on the species present in the image. The following table illustrates a breakdown of tree stem volume based on the total number of trees per DBH:
A stand area of 18 acres was selected in Madison County, N.Y., East of Cazenoia, Moraine Road, Tax Map Number 96-2-1. Low level segmentation was performed to delineate tree crowns, and species classification and tree crown data analysis were performed to determine tree species and total tree stem volume in board-feet. A total of 25,629 board-feet was calculated based on only trees of DBH greater than 14 inches.
Species classification resulted in 45% Hard Maple, 15% Cherry, 4% Red Maple, and 36% undefined. The Norway Spruce Model was selected based on the species present in the image. The following table illustrates a breakdown of tree stem volume based on the total number of trees per DBH:
Timber value was then calculated using the total tree stem volume per species in Doyle and stump prices per 1000 Doyle. The following table illustrates the valuation data generated using the present method:
The foregoing illustrations of embodiments of the methods described herein are offered for the purposes of illustration and not limitation. It will be readily apparent to those skilled in the art that the embodiments described herein may be modified or revised in various ways without departing from the spirit and scope of this disclosure.
What has been described above comprises exemplary embodiments of a remote sensing and probabilistic sampling based forest inventory method. It is, of course, not possible to describe every conceivable combination of components or methodologies for purposes of describing this method, but one of ordinary skill in the art may recognize that many further combinations and permutations are possible in light of the overall teaching of this disclosure. Accordingly, the remote sensing and probabilistic sampling based forest inventory method described herein is intended to be illustrative only, and should be considered to embrace any and all alterations, modifications and/or variations that fall within the spirit and scope of the appended claims. Furthermore, to the extent that the term “includes” may be used in either the detailed description or elsewhere, this term is intended to be inclusive in a manner similar to the term “comprising” as that term is interpreted as a transitional word in a claim.
Claims
1. A remote sensing and probabilistic sampling based method for determining the carbon dioxide volume of a forest, said carbon dioxide volume being a volume of carbon sequestered by said forest, the method comprising:
- processing remote sensing data indicative of tree attribute information for said forest, said remote sensing data comprising at least one of LiDAR data and digital images;
- defining a sampling frame within said remote sensing data;
- determining a field plot corresponding to said sampling frame and collecting field plot data therefrom, said field plot data comprising actual tree attribute information;
- generating a correlated model by combining said field plot data with said remote sensing data corresponding to said sample frame;
- applying said correlated model to all said remote sensing data to produce a probabilistic forest inventory; and
- determining a probabilistic carbon dioxide volume of the forest utilizing said probabilistic forest inventory.
2. The method of claim 1, wherein said correlated model comprises at least one formula for determining at least one of tree species, tree height, tree diameter, and carbon dioxide volume.
3. A remote sensing and probabilistic sampling based method for determining the carbon dioxide volume of a forest, said carbon dioxide volume being a volume of carbon sequestered by said forest, the method comprising:
- processing imagery data, said imagery data indicative of tree attribute information for said forest;
- classifying tree polygons within said imagery data to derive said tree attribute information;
- correlating field data, said field data comprising at least one of actual tree attribute information and plot center location;
- generating a correlated model utilizing said tree attribute information derived from said imagery data and said actual tree attribute information;
- generating a probabilistic forest inventory by applying said correlated model to all said imagery data; and
- determining a probabilistic carbon dioxide volume of the forest utilizing said probabilistic forest inventory.
4. The method of claim 3, wherein generating said correlated model further comprises deriving at least one formula for determining at least one of tree species, tree height, tree diameter, and carbon dioxide volume.
5. The method of claim 3, wherein generating said probabilistic inventory further comprises:
- applying a species probability prediction model, a diameter probability prediction model, and a height probability prediction model to all said imagery data;
- calculating at least a volume per acre; and
- calculating a carbon dioxide volume per acre.
| 5113451 | May 12, 1992 | Chapman et al. |
| 6151424 | November 21, 2000 | Hsu |
| 6864983 | March 8, 2005 | Galle et al. |
| 7062450 | June 13, 2006 | Kobayashi et al. |
| 7412330 | August 12, 2008 | Spicer et al. |
| 7634393 | December 15, 2009 | Tunick et al. |
| 7639842 | December 29, 2009 | Kelle et al. |
| 7884937 | February 8, 2011 | Prasad et al. |
| 20110055220 | March 3, 2011 | Tyburski |
| 0933727 | August 1999 | EP |
- International Search Report for PCT/US2008/087765, Sep. 24, 2009.
- Juha Hyyppa et al: “A Segmentation-Based Method to Retrieve Stem Volume Estimates from 3-D Tree Height Models Produced by Laser Scanners” IEEE Transactions on Geoscience and Remote Sensing, IEEE Service Center, Piscataway, NJ, US, vol. 39, No. 5, May 1, 2001, XP011021749 ISSN: 0196-2892—Abstract, II.B, II.C, IV, V, V.A, VI figure 1.
- Juha Hyyppa et al: “Radar-Derived Standwise Forest Inventory” IEEE Transactions on Geoscience and Remote Sensing, IEEE Service Center, Piscataway, NJ, US, vol. 35, No. 2, Mar. 1, 1997, XP011020825 ISSN: 0196-2892—Abstract and III.A.
- Hyyppa J et al: “Accuracy of forest inventory based on radar-derived stand profile” International Geoscience and Remote Sensing Symposium, 1993. IGARSS '93. Better Understanding of Earth Environment. Tokyo, Japan Aug. 18-21, 1993, New York, NY, USA, IEEE, Aug. 18, 1993, pp. 391-393, XP010114533 ISBN: 978-0-7803-1240-1.
- Hyyppa J: “Feasibility for estimation of single tree characteristics using laser scanner” Geoscience and Remote Sensing Symposium, 2000. Proceedings. IGARSS 2000. IEEE 2000 International Jul. 24-28, 2000, Piscataway, NJ, USA, IEEE, vol. 3, Jul. 24, 2000, pp. 981-983, XP010502852 ISBN: 978-0-7803-6359-5—Abstract and p. 981.
- Hyyppa J et al: “Inventory by compartments using radar-derived stand profiles” Geoscience and Remote Sensing Symposium, 1994. IGARSS '94. Surface and Atmospheric Remote Sensing: Technologies, Data Analysis and Interpretation. International Pasadena, CA, USA Aug. 8-12, 1994, New York, NY, USA, IEEE, US, vol. 1, Aug. 8, 1994, pp. 108-110, XP010138305 ISBN: 978-0-7803-1497-9—Abstract and p. 108.
- International Preliminary Report on Patentability for PCT/US2008/087765, Jun. 22, 2010.
Type: Grant
Filed: Dec 19, 2008
Date of Patent: Feb 7, 2012
Patent Publication Number: 20100040260
Assignee: 2245060 Ontario Ltd. (Hamilton, Ontario)
Inventors: Olavi Kelle (Hatfield, PA), Eric P Macom (Shepherdstown, WV), Robert Pliszka (Sellersville, PA), Neeraj Mathawan (Montgomery, PA)
Primary Examiner: Anand Bhatnagar
Attorney: Spilman Thomas & Battle, PLLC
Application Number: 12/340,217
International Classification: G06K 9/46 (20060101);
